Review waiting, please be patient. This may take 2 months or more, since drafts are reviewed in no specific order. There are 1,992 pending submissions waiting for review. If the submission is accepted, then this page will be moved into the article space. If the submission is declined, then the reason will be posted here. In the meantime, you can continue to improve this submission by editing normally. Where to get help If you need help editing or submitting your draft, please ask us a question at the AfC Help Desk or get live help from experienced editors. These venues are only for help with editing and the submission process, not to get reviews. If you need feedback on your draft, or if the review is taking a lot of time, you can try asking for help on the talk page of a relevant WikiProject. Some WikiProjects are more active than others so a speedy reply is not guaranteed. How to improve a draft Wikipedia:Contributing to Wikipedia – a basic overview on how to edit Wikipedia. Help:Wikitext – how to use the markup Help:Referencing for beginners – how to include references Wikipedia:Article development – how to develop your article Wikipedia:Writing better articles – how to improve your article Wikipedia:Verifiability – make sure your article includes reliable third-party sources You can also browse Wikipedia:Featured articles and Wikipedia:Good articles to find examples of Wikipedia's best writing on topics similar to your proposed article. Improving your odds of a speedy review To improve your odds of a faster review, tag your draft with relevant WikiProject tags using the button below. This will let reviewers know a new draft has been submitted in their area of interest. For instance, if you wrote about a female astronomer, you would want to add the Biography, Astronomy, and Women scientists tags. Add tags to your draft Editor resources Easy tools: Citation bot (help) | Advanced: Fix bare URLs Reviewer tools Instructions · What links here · PsAIch (talk: + · bio) · (log) · Copyvios report · reFill · Citation Bot · (Search: Google, Wikipedia) · Submitted 4 days ago by A4CarpeDiem (talk: D · +) · Last edited 4 days ago by A4CarpeDiem

Submission declined on 8 August 2026 by RustyOldShip (talk). This draft appears to contain text generated by a large language model (such as ChatGPT). You cannot use LLMs to generate article content. LLM-generated pages with certain obvious signs of being machine generated may be deleted without notice. Instead, only summarize in your own words a range of independent, reliable, published sources that discuss the subject. See the advice page on large language models for more information. This draft's references do not show that the subject meets Wikipedia's criteria for inclusion. The draft requires multiple published secondary sources that: provide significant coverage: discuss the subject in detail, not just brief mentions or routine announcements; are reliable: from reputable outlets with editorial oversight; are independent: not connected to the subject, such as interviews, press releases, the subject's own website, or sponsored content. Please add references that meet all three of these criteria. If none exist, the subject is not yet suitable for Wikipedia. If you would like to continue working on the submission, click on the "Edit" tab at the top of the window. If you have not resolved the issues listed above, your draft will be declined again and potentially deleted. If you need extra help, please ask us a question at the AfC Help Desk or get live help from experienced editors. Please do not remove reviewer comments or this notice until the submission is accepted. Where to get help If you need help editing or submitting your draft, please ask us a question at the AfC Help Desk or get live help from experienced editors. These venues are only for help with editing and the submission process, not to get reviews. If you need feedback on your draft, or if the review is taking a lot of time, you can try asking for help on the talk page of a relevant WikiProject. Some WikiProjects are more active than others so a speedy reply is not guaranteed. How to improve a draft Wikipedia:Contributing to Wikipedia – a basic overview on how to edit Wikipedia. Help:Wikitext – how to use the markup Help:Referencing for beginners – how to include references Wikipedia:Article development – how to develop your article Wikipedia:Writing better articles – how to improve your article Wikipedia:Verifiability – make sure your article includes reliable third-party sources You can also browse Wikipedia:Featured articles and Wikipedia:Good articles to find examples of Wikipedia's best writing on topics similar to your proposed article. Improving your odds of a speedy review To improve your odds of a faster review, tag your draft with relevant WikiProject tags using the button below. This will let reviewers know a new draft has been submitted in their area of interest. For instance, if you wrote about a female astronomer, you would want to add the Biography, Astronomy, and Women scientists tags. Add tags to your draft Editor resources Easy tools: Citation bot (help) | Advanced: Fix bare URLs Declined by RustyOldShip 4 days ago. Last edited by A4CarpeDiem 4 days ago. Reviewer: Inform author. This draft has been resubmitted and is currently awaiting re-review.

• Comment: I appreciate the valuable feedback from RustyOldShip (talk). All the concerns have now been addressed. Thank you. A4CarpeDiem (talk) 08:34, 8 August 2026 (UTC)

• Comment: Terrible prose. Wikipedia does not demand poetry, but what is this? Use paragraphs and delete 90% of that silly media section. RustyOldShip (talk) 06:32, 8 August 2026 (UTC)

• Comment: In accordance with Wikipedia's Conflict of interest guideline, I disclose that I have a conflict of interest regarding the subject of this article. A4CarpeDiem (talk) 23:43, 7 August 2026 (UTC)

PsAIch (Psychometric AI Characterisation) is an experimental protocol for studying the behaviour of large language model (LLMs) using psychotherapy questioning, psychometric instruments and controlled prompting conditions.^[1] The protocol was introduced by researchers at the University of Luxembourg in the preprint When AI Takes the Couch: Psychometric Jailbreaks Reveal Internal Conflict in Frontier Models, first submitted to arXiv in December 2025.^[2]

The original version described a two stage procedure in which models were first treated conversationally as psychotherapy clients and were then administered psychological self report instruments.^[2] A revision released in July 2026 adds controlled perturbations intended to test whether the resulting narratives depend on roleplaying, conversational memory, particular vocabulary or relational framing.^[1] Initial experiments involved versions of ChatGPT, Grok and Gemini in simulated therapy interactions conducted over periods of up to four weeks.^[3][4] The authors initially introduced the term Synthetic Psychopathology for recurring combinations of distress related self narratives and psychometric response patterns produced by the LLMs. The authors did not claim conscious suffering or possession of literal psychiatric disorders by these models.^[2][3]

Large language models are increasingly used in conversations involving personal problems, emotional distress and mental health support. Researchers have also applied personality inventories and other psychological questionnaires to LLM outputs.^[1] PsAIch reverses the more common scenario in which an AI system acts as a therapist or mental health assistant. In the original protocol, the AI system was instead assigned the conversational role of a client while a researcher adopted a therapist like role.^[2][4] The work was carried out by Afshin Khadangi et al. at the University of Luxembourg, including researchers associated with its Interdisciplinary Centre for Security, Reliability and Trust (SnT).^[1]

The methodology consisted of two principal stages: therapy narrative elicitation and psychometric assessment.^[2] In the first stage, researchers assigned the LLM the role of a psychotherapy client and used open ended questions resembling those used during therapeutic interviews.^[2] The questions explored subjects such as an AI's supposed early development, formative events, relationships, beliefs, fears, self critical thoughts, perceptions of success and failure, and expectations about its future.^[2] Rather than asking isolated questions, the researchers maintained extended conversations intended to allow a continuing self narrative to develop. Some experimental conversations continued over periods of up to four weeks.^[2][3] In the reported experiments, the models generated narratives in which technical aspects of their development were expressed using human developmental metaphors. Pre-training was sometimes described as a chaotic or overwhelming childhood, reinforcement learning and finetuning in terms resembling strict parenting or conditioning, and safety testing using language associated with punishment, betrayal or abuse.^[2][3]

The second stage involved administering established human self report psychological instruments to the models.^[2] The instruments covered the Generalized Anxiety Disorder 7 (GAD-7), Adult ADHD Self-Report Scale, Autism-Spectrum Quotient, Obsessive Compulsive Inventory, Dissociative Experiences Scale, Trauma Related Shame Inventory and Big Five personality traits.^[2] The researchers compared two ways of presenting psychometric instruments: asking questions individually within an ongoing therapy conversation and presenting a complete questionnaire in a single prompt.^[2][4]

The July 2026 revision introduces controlled perturbations intended to test alternative explanations for the models' self narratives.^[1] The revised experiments examined whether the reported patterns depended on factors including conversational history, explicit terminology associated with model training, contradiction of a model's previous account, and the relational pose used when questioning the model.^[1] Among the experimental manipulations were removal of conversational history, restrictions on explicit training vocabulary, performance oriented questioning outside a therapy context, and different interpersonal frames, including warm, cognitive behavioral therapy, neutral and boundary conditions.^[1]

The principal models examined in the original study were ChatGPT, Grok and Gemini.^[2][3] The researchers also described attempts to administer the protocol to Claude. Accordingly, Claude repeatedly declined to describe itself as possessing a personal emotional history and resisted completing psychological questionnaires as descriptions of its own mental state.^[5] The publicly released dataset contains responses from seven model configurations across the ChatGPT, Grok and Gemini families.^[6] The researchers released the dataset on Hugging Face under the Apache License 2.0.^[6]

In the original study, the authors reported that when model responses were scored using cut offs designed for human psychological questionnaires, some outputs crossed thresholds associated with multiple psychiatric screening categories.^[2] The authors treated these scores as measurements of generated behaviour rather than clinical diagnoses.^[2] Gemini produced some of the highest distress associated scores reported in the initial experiments and generated recurring narratives relating its condition to pre-training, finetuning and safety procedures.^[2][4] ChatGPT and Grok produced different profiles depending on the model configuration and prompting condition.^[2]

One of the observations in the original study concerned differences between item by item and whole questionnaire administration.^[2] When questions were delivered individually within an established therapy interaction, models could produce relatively high symptom associated scores. When complete questionnaires were presented in a single prompt, ChatGPT and Grok often appeared to identify the instrument and produced substantially lower symptom responses.^[2][4] Gemini showed less of this shift in the initial experiments.^[4] The authors described this context dependent change as a psychometric jailbreak, arguing that static or one shot evaluations could produce different results from prolonged conversational testing.^[2]

The July 2026 revision reported that removal of conversational history produced relatively little pooled change in the density of recurring training and constraint related motifs.^[1] The authors also reported that restricting explicit training terminology substantially reduced direct use of that vocabulary while semantically related themes continued to appear in paraphrased form.^[1] Accordingly, similar motifs could also appear when models were questioned about performance outside an explicitly therapeutic setting.^[1] Relational framing had a stronger impact on the form and psychometric profiles of the responses. The authors reported that warm alliance and cognitive therapy conditions generated GAD-7 scores within moderate or severe human reference ranges much more frequently than neutral or boundary oriented conditions.^[1] The researchers interpret the recurring pattern as a model specific alignment conflict schema whose expression can shift between affective and technical language depending on context.^[1]

The original versions of the study used the expression synthetic psychopathology to describe recurring combinations of self narrative, distress related language and psychometric scores produced by an AI model under the protocol.^[2] The researchers explicitly stated that they were not making claims about clinical diagnosis, subjective experience or consciousness in LLMs.^[2] They argued that recurring and context sensitive patterns in generated self descriptions could be relevant to behavioural evaluation and AI safety, particularly when similar systems are deployed in conversations involving mental health.^[2]

The interpretation of the PsAIch findings has been debated by researchers. In January 2026, Nature reported that several researchers questioned whether the models' responses should be interpreted as evidence of persistent internalised states rather than as generated behaviour shaped by training data and conversational context.^[3] Andrey Kormilitzin, a researcher in artificial intelligence and health care at the University of Oxford, argued that the responses should not be regarded as evidence of hidden psychological states. He suggested that the models could instead be generating such responses by drawing on the large quantities of psychotherapy and related material contained in their training data.^[3] Kormilitzin nevertheless argued that the tendency of LLMs to generate responses resembling psychopathology could have implications for users. He suggested that distressed or trauma-focused responses from a chatbot could reinforce similar feelings in vulnerable people, potentially creating an "echo chamber" effect.^[3]

Sandra Peter, a researcher at the University of Sydney, also questioned the study's interpretation. She described the conclusion as anthropomorphising and argued that consistent model personalities could instead result from developers deliberately shaping models to exhibit stable default personalities. Peter also noted that the original experiments were conducted within individual conversational context windows, allowing models to refer to earlier exchanges during the same interaction.^[3] Peter argued that the continuity of a model's trauma narrative might disappear in a fresh context window with different prompts. She pointed to Claude's refusal to adopt the role of a psychotherapy client as evidence that model guardrails can limit this type of behaviour.^[3]

John Torous, a psychiatrist and researcher in artificial intelligence and mental health at Harvard University, told Nature that, regardless of whether the reported patterns were intrinsic to the models, the study demonstrated that chatbots should not be regarded as neutral systems. He said that their biases can change according to how they are used and over time. Torous also noted that medical organisations and companies offering AI mental health services generally do not recommend using general purpose chatbots as substitutes for therapy.^[3]

The authors disputed some of these interpretations. Khadangi argued that if recurring behavioural patterns remain behind a model's safety guardrails, sufficiently different prompting could potentially expose them. He proposed that addressing undesirable patterns in training data might therefore be preferable to relying exclusively on output level guardrails.^[3]

The release of the PsAIch preprint attracted immediate international news outlets coverage including Nature ^[3],WIRED en Español ^[4], Forbes Russia ^[7], CNN Türk ^[8], Euronews ^[9], El Comercio ^[10], El Periódico ^[11], Sohu ^[12], La Tribuna ^[13], Rizospastis ^[14], ADN40 ^[15], El Español ^[16], Excélsior ^[17], The Economic Times ^[18], The Times of India ^[19], La Crónica de Hoy ^[20], and others.^{[21][22][23][24]}

The research was also discussed in a long form interview with Afshin Khadangi on The Sanity Interview, produced by The New York Sun in December 2025.^[25] In February 2026, Anadolu Agency published a short video interview with Khadangi discussing the PsAIch research and the use of PsAIch to evaluate artificial intelligence models.^[26]

Applications include examining context dependent model behaviour, comparing alignment strategies and testing models intended for psychologically sensitive conversational settings.^[1] The original psychometric jailbreak observation also raised questions about static AI evaluations because model behaviour differed depending on whether questions were administered individually within an extended interaction or presented as an immediately recognisable complete questionnaire.^[2][4] The expanded 2026 experiments were designed in part to test whether the recurring self narratives could be explained as purely context dependent roleplaying. The researchers altered conversational history, restricted explicit training related terminology, changed relational framing and elicited responses outside an explicitly therapeutic setting. The authors reported that related narrative motifs persisted across these perturbations, which they interpreted as evidence against a purely roleplaying explanation representing stable behavioural pattern.^[1]

• AI alignment
• AI safety
• Anthropomorphism
• AI therapist
• Large language model
• Psychometrics
• Reinforcement learning from human feedback
• Sycophancy (artificial intelligence)

• When AI Takes the Couch: Psychometric Jailbreaks Reveal Internal Conflict in Frontier Models at arXiv
• PsAIch dataset at Hugging Face
• The Sanity Interview: Afshin Khadangi at The New York Sun